Voice synthesis GPU peak allocated — C013
Voice candidates were scored blind by embedding similarity and intelligibility, then chosen by listening — samples are not published.
Single reading
1,990 MiB
Protocol
Test definition
Offline ECAPA-TDNN speaker-embedding cosine to a fixed TEST-only reference centroid and Whisper round-trip WER over eight fixed original sentences unless stated. Cold model load 10.77 s.
Limits
Limitations
- MEASURED_UNQUALIFIED. Similarity is diagnostic, not identity or quality; scores are only comparable within one reference population. Owner preference did not track similarity. DO NOT publish audio (rights flag in voice.md).
Log
Recorded values
| Date | Value | Evidence | Samples | Source |
|---|---|---|---|---|
| 13 September 2026 | 1,990 MiB | MEASURED | 8 | — |
Source paths refer to the private JARVIS repository and runtime records. They are listed for audit and are not published.